Tweedie loss is a candidate for count-data problems where targets combine many zeros with otherwise continuous, nonnegative values. It offers a distribution-based alternative to training separate quantile models when the predictive distribution itself matters. section 5.3.1
Its appeal comes from covering several target behaviors:
- Includes continuous distributions such as normal, gamma, and inverse Gaussian. section 5.3.1
- Includes discrete scaled Poisson and compound Poisson–gamma distributions. section 5.3.1
- Models positive mass at zero with continuous values elsewhere, matching the stated count-data pattern. section 5.3.1
Retail demand experiment
A retail forecasting team can review Tweedie loss as a second experiment beside six quantile-loss models, comparing its suitability for count data against quantile-specific guarantees. section 5.3.1
